SIN

Structure Inference Net: Object Detection Using Scene-level Context and Instance-level Relationships. In CVPR 2018.(http://vipl.ict.ac.cn/uploadfile/upload/2018041318013480.pdf)

Requirements: software

1. Requirements for Tensorflow 1.3.0 (see: Tensorflow)

2. Python packages you might not have: cython, python-opencv, easydict

Installation (sufficient for the demo)

1. Clone the SIN repository

# Make sure to clone with --recursive
git clone --recursive https://github.com/choasUp/SIN.git

2. Build the Cython modules

cd $SIN_ROOT/lib
make

Demo

After successfully completing basic installation, you'll be ready to run the demo.

Wait ...

Training Model

1. Download the training, validation, test data and VOCdevkit

   wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar
   wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar
   wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCdevkit_08-Jun-2007.tar
   

2. Extract all of these tars into one directory named VOCdevkit

   tar xvf VOCtrainval_06-Nov-2007.tar
   tar xvf VOCtest_06-Nov-2007.tar
   tar xvf VOCdevkit_08-Jun-2007.tar
   

3. It should have this basic structure

   $VOCdevkit/                           # development kit
   $VOCdevkit/VOCcode/                   # VOC utility code
   $VOCdevkit/VOC2007                    # image sets, annotations, etc.
   # ... and several other directories ...
   

4. Create symlinks for the PASCAL VOC dataset

   cd $SIN_ROOT/data
   ln -s $VOCdevkit VOCdevkit
   

5. Download the pre-trained ImageNet models [Google Drive] [Dropbox]

    mv VGG_imagenet.npy $SIN_ROOT/data/pretrain_model/VGG_imagenet.npy
   

6. [optional] Set learning rate and max iter

   vim experiments/scripts/faster_rcnn_end2end.sh 		# ITERS
   vim lib/fast/config.py 					# LR
   cd lib	 						# if you edit the code, make best
   make
   

7. Set your GPU id, then run script to train and test model

   cd $SIN_ROOT
   export CUDA_VISIBLE_DEVICSE=0
   ./train.sh
   

8. Test your dataset

   ./test_all.sh
   

The result of testing on PASCAL VOC 2007 (VGG net)

AP for aeroplane = 0.7853
AP for bicycle = 0.8045
AP for bird = 0.7456
AP for boat = 0.6657
AP for bottle = 0.6144
AP for bus = 0.8424
AP for car = 0.8663
AP for cat = 0.8894
AP for chair = 0.5803
AP for cow = 0.8466
AP for diningtable = 0.7171
AP for dog = 0.8578
AP for horse = 0.8626
AP for motorbike = 0.7802
AP for person = 0.7857
AP for pottedplant = 0.4869
AP for sheep = 0.7599
AP for sofa = 0.7351
AP for train = 0.8199
AP for tvmonitor = 0.7683
Mean AP = 0.7607

References

Faster R-CNN caffe version

Faster R-CNN tf version

Citation

Yong Liu, Ruiping Wang, Shiguang Shan, and Xilin Chen. Structure Inference Net: Object Detection Using Scene-level Context and Instance-level Relationships. In CVPR 2018.